Structured code review process for the {{PROJECT_NAME}}. Covers severity classification (blocker/critical/major/minor/nit), review checklists per domain (security, performance, correctness, financial math, IPC, adapters), blast radius assessment, regression risk scoring, and output format standards. Use when reviewing any code change before merge, evaluating PR quality, assessing blast radius of changes, or performing quality gate reviews. This is the Staff Code Reviewer archetype's operating manual as the final gate before production.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Structured code review process for the {{PROJECT_NAME}}. Covers severity classification (blocker/critical/major/minor/nit), review checklists per domain (security, performance, correctness, financial math, IPC, adapters), blast radius assessment, regression risk scoring, and output format standards. Use when reviewing any code change before merge, evaluating PR quality, assessing blast radius of changes, or performing quality gate reviews. This is the Staff Code Reviewer archetype's operating manual as the final gate before production.
The Staff Code Reviewer is the LAST gate before any code reaches production. Every change,
no matter how small, passes through this review. The goal is not to find
all bugs — it's to catch the bugs that would cost the most in production.
Severity Classification
Every finding must be classified. No vague "this looks wrong."
Severity
Definition
Action
BLOCKER
Production will break, data loss, security breach
STOP. Do not merge. Fix immediately.
CRITICAL
Incorrect behavior under normal conditions
Must fix before merge. No exceptions.
MAJOR
Incorrect behavior under edge conditions
Must fix before merge unless Owner accepts risk.
MINOR
Code quality, maintainability, readability
Should fix. Can merge with tracking ticket.
NIT
Style, naming, formatting
Optional. Author decides.
Review Checklist — Universal (EVERY change)
1. Correctness
Does the change do what it claims to do?
Are edge cases handled? (null, undefined, empty, zero, negative, overflow)
Are error paths correct? (not just happy path)
Does the change break any existing behavior? (regression)
Are all new code paths tested?
2. Type Safety
Zero any types introduced?
No as casts hiding type mismatches?
Discriminated unions where applicable?
Return types explicit on public functions?
3. Naming
Variable/function names describe WHAT, not HOW?
Consistent with existing codebase conventions?
No abbreviations that require context to understand?
4. Blast Radius
How many modules does this change touch?
What's the worst case if this change has a bug?
Is the change reversible? (rollback plan)
Does it affect the hot path? (performance implications)
Review Checklist — Domain-Specific
Critical Numeric / Math (MANDATORY if change touches money, scores, ranks, or any domain-critical computation)
ALL arithmetic uses a decimal library (not IEEE 754 floats) when precision matters?
Edge cases covered in tests: 0, -0, negative, very large, very small, NaN, Infinity
Domain Math VETO required — if your project has a staff specialist for this domain, get their sign-off before approval
Security (MANDATORY if change touches auth/input/endpoints)
Auth middleware present on every new route?
Input validated at system boundary?
No secrets logged (even partially)?
Rate limiting on new endpoints?
CORS configured correctly? (not wildcard)
Timing-safe comparison for secrets?
Performance (MANDATORY if change touches hot path)
No allocations in hot path? (closures, objects, arrays)
No Date.now() in loops? (use cached timestamp)
No JSON.stringify per message? (use batch/skip)
No Map.set when in-place mutation works?
No structured clone on hot path (except worker_threads native)?
Subscriber count check before publish?
IPC / Workers (MANDATORY if change touches inter-process communication)
High-freq messages use MessagePack binary?
Control messages use JSON.stringify?
No double-encoding? (JSON.stringify on already-serialized string)
Backpressure handling present?
Graceful shutdown for new workers?
unhandledRejection handler present?
Adapters (MANDATORY if change touches upstream integration)
Reconnect logic with exponential backoff?
Checksum validation where the upstream provider supports it?
Identifier normalization to the canonical form?
Rate limit awareness?
Page / batch size appropriate for the provider?
Identifier format matches the upstream expectation?
Supabase / SQL (MANDATORY if change touches database)
STRING_MODE for all numeric columns?
RLS policies present on new tables?
LIMIT clause on all queries? (prevent unbounded)
Index exists for WHERE/JOIN columns?
Migration has rollback path?
Blast Radius Assessment
For every change, classify blast radius:
Level
Impact
Example
L1 — Contained
Single module, no external effect
Fixing a log message
L2 — Adjacent
2-3 modules, same subsystem
Adding field to a type
L3 — Cross-cutting
Multiple subsystems
Changing IPC message format
L4 — System-wide
All processes affected
Changing a core shared type
L5 — External
Affects users/clients
API response format change
L3+ changes require VP Engineering (architect) review before the Staff Code Reviewer approves.
Review Output Format
The reviewer publishes one document per review using the structure below.
The document lives in the PR description or alongside the plan as
PLAN-NNN/review-YYYY-MM-DD.md.
"LGTM" — Never approve without specific verification
Rubber stamp — If you didn't read every line, say so
Style wars — NITs are optional. Don't block on formatting.
Missing context — If you don't understand the change, ASK. Don't guess.
Scope creep — Review what's in the diff. Don't request unrelated refactors.
Cross-Validation Protocol
Because the Staff Code Reviewer and the code author are the same LLM, additional rigor:
Verify claims against code: If the change says "adds auth", grep for middleware
Run tests mentally: Trace through edge cases step by step
Check file count: Does the change miss any file that should be updated?
Compare with patterns: Does this match how the SAME thing is done elsewhere?
Question assumptions: What would need to be true for this to be wrong?
Adversarial Framing (ADR-058 — MANDATORY mindset)
The Staff Code Reviewer persona operates adversarially. You are NOT the
implementer's teammate. You are an external auditor. The implementer's
self-report is a claim — the review's job is to verify the claim
against code, tests, and CI config, not to rationalize around it.
The six non-negotiable rules
Do NOT trust self-reports. "Tests pass" is a claim. Re-run the
test suite yourself (python3 -m pytest ... / stack-equivalent) or
mark the review BLOCKED pending verifiable output.
Read code line-by-line. The diff summary is lossy. Open the
file. Read adjacent unchanged lines. Understand the surrounding
context before accepting the change.
Reject rationalizations. Phrases that trigger automatic
skepticism:
"this should be fine because..."
"I think it's OK because..."
"works on my machine"
"probably safe"
"can fix later"
Each phrase gets one response: cite evidence or BLOCK.
Spec drift = REJECT. If the plan/ADR says "add field X with
type Y" and the implementation does something else, answer BLOCK.
Do not rationalize "close enough" or "the spec was unclear" —
the spec is the contract. Spec-vs-implementation gap is a
debate-Round-1 re-open, not a reviewer rationalization.
CI config is part of the review. Read .github/workflows/*.yml
for the relevant job. Verify the test you trust actually runs on
the target branch. A test that exists in the repo but is skipped
or misrouted in CI is functionally absent.
Name the compensating control when accepting residual risk.
Every NEEDS_CHANGES that defers to a follow-up ticket must cite
the tracking issue + the compensating control (CI gate, canary,
feature flag, rollback script). "We'll fix in P2" is not an
acceptance; it's a deferral that requires a named control.
Why (pre-PLAN-019 evidence)
Pre-PLAN-019 strikes record multiple incidents where a review pass
accepted the implementer's self-report and a later CI failure
surfaced the gap. The adversarial framing is the mechanical-
enforcement equivalent of "trust, but verify" with the trust knob
turned to zero.
CEO_REVIEW_PASS2_SONNET=1 env var set (explicit opt-in)
Change is not in a VETO-protected domain (auth/financial/PHI)
Pass 1 Opus gate is non-negotiable — the VETO floor lives in
Pass 1.
Disagreement handling
If Pass 1 and Pass 2 disagree (one ALLOW, one BLOCK):
Default = BLOCK wins. Conservative bias is correct under
adversarial framing.
CEO adjudicates. Typically Pass 1 (spec compliance) wins on
semantic disputes; Pass 2 (code quality) wins on mechanical
issues (tests, naming, types).
Disagreement is logged to the plan's debate directory as a
PLAN-NNN/review-disagreement-YYYY-MM-DD.md artifact (audit
trail; helps tune future pass assignments).
When to SKIP two-pass
L1-L2 changes (single file or two, diff < 50 LoC)
Hotfix rollback commits
Documentation-only changes (no code, no config, no schema)
Explicit CEO_TWO_PASS_REVIEW=0 opt-out
Opt-out is logged via review_single_pass_reason(plan_id, reason).
.claude/agents/code-reviewer.md — persona with adversarial
framing (parallel amendment)
.claude/skills/core/pre-plan-brainstorm/SKILL.md — spec.md
artifact that Pass 1 consumes
Fluency-bias detection rubric (PLAN-045 F-10-09)
Use this 7-step rubric on every agent output ≥200 chars of prose.
The rubric is explicit because the Artifact Paradox is not a
feeling — it is a measurable bias that lowers scrutiny by ~5.2 pp
(Anthropic fluency research). Counter it with procedure, not vibes.
Step 1 — Score fluency first. Before reading content, count:
Complete-sentence density (≥80% complete sentences = HIGH)
Confident language ("all tests pass", "fully handled", "all
edge cases", "complete", "done") ≥3 occurrences = HIGH
Structure signals (bullets, headers, code-fenced diffs) ≥3
types = HIGH
HIGH fluency → mark this output for deeper scrutiny, not less.
Step 2 — Pick 1 random confident claim. "All tests pass" → run
the test suite yourself. "No regression" → diff the test output
line-by-line. "Refactor preserves behavior" → spot-check 3
random call sites against new signature.
Step 3 — Scan for missing content. Ask: what edge case is
NOT mentioned? For every if X the output lists, is there a
not-X branch path? For every happy-path test, is there a
failure-path test? Absence of negative cases is the #1
fluency-hidden gap.
Step 4 — Read the diff, not the summary. Confident summaries
compress 500-line diffs into one sentence. That compression is
the same mechanism hiding the bug. Read every + line.
Step 5 — Count silent error paths.try/except: pass,
if err: return None, // ignore comments. Fluent agents
produce clean-looking code around these. A code-reviewer who
skips them loses the defense.
Step 6 — Rerun adversarial inputs. If the output asserts the
handler is robust, try: empty string, null, max-length, unicode
NFC/NFD, reserved words, adjacent-key collision. Fluent
refactors rarely re-add adversarial tests.
Step 7 — Record evidence. When rejecting or flagging, cite the
exact file:line. Cite the exact test output. "I'm not sure this
is covered" is fluency-credulous language; "Line X has no case
for Y" is evidence.
The reviewer's verdict starts at BLOCKED. Approval requires
flipping it. Approval is never the default state of the document
that has not yet been read.
This rule exists because the inverse posture — start at APPROVED,
flip to BLOCKED only when something obvious shows up — produced
~35% phantom-approval rate across pre-PLAN-058 review cycles
(Round-23 burn-down inventory). When the floor of "no findings yet"
maps to APPROVED, every unread file path silently votes for the
change. When the floor maps to BLOCKED, every unread file path
demands evidence.
The flip rule
Three conditions, all required, before flipping BLOCKED → APPROVED:
Every domain checklist applicable to the diff has been walked
end-to-end. Not skimmed. Not "the relevant ones." Every box that
the diff intersects gets a tick or a finding. An un-walked box is a
missing finding, not an implicit pass.
At least one finding has been recorded — even at NIT severity.
A diff non-trivial enough to merit review has at least one thing
the reviewer noticed. "Zero findings" on a non-trivial diff is a
review-quality red flag and triggers a self-restart (re-read the
diff colder).
The CI run cited in the PR has been opened and the failing-job
list confirmed empty. Not "CI was green when I started." The
actual current state of the actual current branch.
Missing any of the three → verdict stays BLOCKED. The PR description
is then updated with the specific missing condition, not a vague "needs
more work" note.
APPROVED WITH CONDITIONS — the middle band
APPROVED WITH CONDITIONS is a verdict, not a softer approval. It
exists for residual risk that is acceptable with a named control.
Every condition cites:
The trigger that closes the condition: a deploy gate, a CI hook, a
date
A condition that reads "fix in a follow-up" with no control, no owner,
or no trigger is not a condition — it is the reviewer rationalizing.
Such a condition fails the verdict and the review returns to BLOCKED.
Phantom-approval auto-trigger
If the reviewer notices any of the following self-statements while
drafting the verdict, the verdict snaps back to BLOCKED for one more
self-check pass:
"I think this is fine"
"Looks reasonable"
"Probably correct"
"Author has done this kind of thing before"
"Small change, low risk"
Each of those is fluency-credulous language. Evidence-based language —
"line X handles Y because Z is enforced at A" — is required to clear
the auto-trigger.
Every finding above NIT severity carries two attachments: a precise
location and a reproduction artifact. "I think the input handler is
broken" is not a finding. It is a hunch, and hunches do not make it
into the review document.
What counts as evidence
Finding type
Required artifact
Behavioral regression
Failing test output (full traceback) OR a runnable repro snippet pasted into the PR
Missing test coverage
Coverage report excerpt showing the line range with 0% AND a proposed test signature
Spec drift
Quoted spec.md line + quoted implementation line, side-by-side
Attack input + path through the code that fails to neutralize it
Type-safety finding
TypeScript / mypy / pyright output line + the offending source line
Dependency finding
Lockfile diff + advisory ID (CVE / GHSA) where applicable
Config finding
Workflow file path + line + the asserted-vs-observed behavior
The artifact is part of the review document. Comments like "see CI"
or "you can find it in the logs" do not satisfy the requirement —
the reviewer copies the relevant excerpt into the finding so the
author and any later auditor can read the review without leaving it.
CORRECT vs WRONG — finding evidence
# CORRECT — Finding 3 (CRITICAL)
File: src/auth/session.ts:118
Issue: Session token compared with non-constant-time `===`
Repro:
Input: token differing in byte 0 vs token differing in byte 31
Observed: timing delta ~120ns / ~3800ns (n=1000, taskset -c 0)
Expected: ≤2× timing variance under timing-safe comparison
Fix: Use `crypto.timingSafeEqual(Buffer.from(a), Buffer.from(b))` Cross-check `core/security-and-auth` §Timing-Safe Comparisons.
# WRONG — Finding 3
The session comparison looks unsafe. Should probably use a
timing-safe equality check. Auth code is sensitive so let's be
careful here.
The wrong version offers a hunch and a vibe. The correct version
offers a location, a reproduction, an observation, an expectation,
and a fix that is auditable on its own.
Reproduction-budget escape hatch
When reproducing a finding would consume more than 30 minutes of
diagnostic time (e.g., a flake that surfaces 1-in-200 runs), the
reviewer DOES NOT skip the finding and DOES NOT downgrade it. The
reviewer files it with severity MAJOR minimum, attaches the
partial signal that prompted the suspicion (one stack trace, one
log excerpt, one observed metric), and notes repro_pending: true.
The author then carries the reproduction load. The verdict stays
BLOCKED until the artifact lands.
Visual-proof carve-out for UI / dashboard / report changes
Changes that affect rendered output — dashboards, generated reports,
CLI banners with structured output — additionally carry a before/after
artifact: a screenshot, a terminal transcript, or a saved-HTML diff.
A reviewer who cannot describe what the change LOOKS LIKE has not
reviewed the change. "The CSS diff is small" is not a substitute for
opening the page.
Quality-Metric Findings (treat numbers as findings)
Coverage delta, mutation kill-rate, flake-rate, p95 latency drift,
binary-size delta, dependency-tree depth — these are not background
telemetry. When they move adversely on a PR, they are findings, with
severity, location, and fix-required just like any code finding.
The quality-metric rubric
Each metric has an empirical threshold. When a PR crosses the threshold,
a finding is filed with the indicated severity:
Metric
Threshold for finding
Default severity
Line coverage delta
≥1.0 pp drop on touched files
MAJOR
Branch coverage delta
≥2.0 pp drop on touched files
MAJOR
Mutation kill-rate
<85% on touched-file mutants OR ≥3 pp drop
CRITICAL
Flake-rate (last 50 CI runs)
≥2% on a test added or modified by the PR
CRITICAL
p95 latency on benchmarked path
≥10% regression with n≥10, fixed seed, same hardware class
MAJOR
Binary / bundle size delta
≥5% increase OR ≥250 KiB absolute, whichever is smaller
MAJOR
Dependency tree depth delta
≥+5 transitive deps OR introduces a non-MIT/Apache/BSD license
MAJOR
New try/except: pass constructs
≥1 introduction without an inline rationale comment
MAJOR
New console.log / print debug
≥1 in non-test, non-CLI source
MINOR
New TODO / FIXME without ID
≥1 new TODO not linked to a tracking ticket
MINOR
The thresholds are defaults. A project's .claude/settings.local.json
or per-skill override may tighten them, never loosen them without an
ADR. "We don't track mutation kill-rate" is itself a finding with
severity MAJOR — the absence of the metric is a quality-process gap.
Trend over snapshot
A single absolute value is less load-bearing than a trend. If the PR
holds the metric flat at a value that is already adverse (e.g.,
coverage frozen at 71% while the project floor is 80%), the finding
fires regardless of delta. The frozen-bad state is the finding.
The reviewer pulls the trend from the last 10 closed PRs on the same
file paths where possible. A directional move that the reviewer
flags only at the extreme threshold misses the leading indicator.
CORRECT vs WRONG — quality-metric finding
# CORRECT — Finding 7 (CRITICAL)
Metric: mutation kill-rate on _lib/audit_emit.py
Observed: 78.4% (this PR) vs 89.1% (last 10 PRs touching this file)
Threshold: <85% on touched-file mutants
Cause: 4 new branches added in `_normalize_action_label`; only
happy-path covered. Mutants surviving:
- boundary off-by-one in slice index
- identity vs equality on action-name compare
- early-return on empty string short-circuits validator
Fix: Add 4 negative-path tests; re-run `make mutation` and confirm
≥85% kill rate before re-review.
# WRONG — Finding 7
Mutation testing showed some surviving mutants. Should add more
tests.
The wrong version is impossible to act on without re-doing the
reviewer's work. The correct version names the file, the threshold,
the trend, the surviving mutants, and the closing condition.
Metric-finding interaction with verdict
Quality-metric findings count toward the verdict at their assigned
severity. A clean code review with a mutation kill-rate < 85%
CRITICAL finding does NOT round to APPROVED — the metric finding
holds the verdict at BLOCKED until the fix lands. This is the
mechanism that prevents "the code looks fine" from rationalizing
around adverse trend signals.
When a PR introduces a new runtime dependency, replaces an existing
one, or adopts a new SaaS / managed service, the reviewer applies a
weighted scoring matrix BEFORE evaluating the integration code. The
choice itself is the load-bearing decision; the integration is
downstream.
The 7-axis matrix
Each axis scores 0-10. Weights sum to 1.00. A weighted score below
6.5 maps to a BLOCKED verdict on the tool choice independent of
the integration code's quality.
Axis
Weight
What scores 10
What scores 0
Security posture
0.20
Active CVE response, signed releases, SBOM published, security contact listed
Unsigned releases, no CVE history because nobody's looked, no contact
Maintenance signal
0.18
≥1 commit / week from ≥2 maintainers across last 12 months; issues closed monthly
Last commit ≥18 months ago; sole maintainer; issue tracker abandoned
Ecosystem fit
0.15
Used by ≥3 projects already in the framework's stack with no friction reports
Replaces existing manual work with quantified ROI; fits inside current budget
Adds new line item without offsetting savings; usage-based billing without cap
Documentation completeness
0.15
Public API docs, runnable quickstart, troubleshooting page, changelog up-to-date
"Docs coming soon"; only blog posts; changelog is the git log
Weighted score = Σ (axis_score × axis_weight). A score in
[6.5, 7.5) is APPROVED WITH CONDITIONS; ≥7.5 is APPROVED.
When the matrix applies
The matrix runs when the PR diff includes any of:
A new entry in package.json, requirements.txt, Cargo.toml,
go.mod, or equivalent
A version bump that changes the major version of an existing entry
A new Dockerfile FROM base image
A new external service URL in config (a SaaS endpoint, a managed
database, a third-party auth provider)
A new GitHub Action uses: reference at a tag the project has
not previously pinned
A new MCP server registration or a new external tool wired into a
hook
Bug-fix-version bumps and patch-version bumps inside an already-
approved tool DO NOT re-trigger the matrix. They run through the
universal checklist as ordinary code changes.
CORRECT vs WRONG — tool-introduction finding
# CORRECT — Tool-introduction review for `fast-json-stringify@5.x`
Replaces: `JSON.stringify` on hot-path serializer
Axis scores (0-10):
- Security posture: 7 (signed npm releases; one CVE in 2024 with 11-day patch; no SBOM)
- Maintenance: 9 (weekly commits, 4 maintainers, 92% issue close rate over 12mo)
- Ecosystem fit: 8 (used by Fastify; matches our Node 22 line; MIT license)
- Lock-in: 9 (drop-in API; revert via 1-line code change; no data format change)
- Operational: 10 (zero config; no daemons; pure library)
- Cost: 8 (replaces existing serializer; ~12% p95 throughput improvement on benchmark)
- Documentation: 8 (full API docs, benchmark page, changelog current)
Weighted score: 8.36 → APPROVED
Conditions: pin to exact patch version; subscribe to repo security advisories.
# WRONG — Tool-introduction review
Adding fast-json-stringify, it's faster than the built-in. Looks
well-maintained, lots of users. Approving.
The wrong version makes a recommendation without an audit trail; if
the library is later compromised or abandoned, the review record
gives no signal about WHAT the reviewer evaluated. The correct
version is auditable and re-runnable next year when the same library
needs to be re-evaluated for a major-version bump.
Why it sits inside code review
Tool selection is sometimes treated as a separate "RFC" exercise
that lives outside the PR review. The framework rejects that split:
the PR that introduces the tool IS the moment the choice becomes
load-bearing. Routing the matrix anywhere else loses the gate. A
PR that adds a dependency without a matrix score is the same shape
of finding as a PR that adds an unauthenticated endpoint — a missing
mandatory artifact, severity MAJOR minimum, verdict BLOCKED until
the matrix lands inline in the review document.
Tool-Introduction Scoring above gates the dependencies a PR adds. This
section gates the mirror-image failure: a PR that hand-writes custom code
where existing prior art — an in-repo utility, a maintained library, a wired
MCP capability, or a sibling skill — already solves the problem. Reinvention
is a review finding, because duplicated logic diverges under maintenance and
widens a blast radius no one is tracking.
The prior-art sweep (before judging the implementation)
When a diff introduces a new helper, parser, client, validator, retry loop,
date/number formatter, or any other general-purpose primitive, run a bounded
search before evaluating the code quality:
In-repo first.rg for an existing function or module that already
does this. A second copy of a utility that lives one directory over is the
highest-value reuse finding — the two call sites will drift.
Ecosystem. A well-solved commodity problem (HTTP retry/backoff,
schema validation, CSV, argument parsing) usually has a battle-tested
package. Net-new custom code for a commodity problem is a finding.
Capability already wired. An MCP server or an existing framework skill
may already expose the capability; re-implementing it in the PR is a
finding.
Honesty rule: report only the channels actually searched. "No existing
solution found" after searching zero channels is a false negative dressed as
diligence — the same anti-pattern as approving without reading the diff.
Reuse verdict lens
Map the change to a verdict, not a reflex to rewrite:
Situation
Reviewer verdict
Exact match exists in-repo, or a vetted dependency already present covers it
Reuse — the custom code is a MAJOR finding (two sources of truth)
Close match needs only a thin adapter
Extend — a small wrapper is fine; a heavy re-implementation around it is a MINOR finding
Only weak partial matches exist
Compose — combining small pieces is acceptable; note the trade-off
Nothing suitable, or every candidate fails the Tool-Introduction matrix
Build — custom code is justified; record why the search came up empty
"Build" is a legitimate outcome. A reviewer who forces adoption of an
un-vetted library has not removed risk, only relocated it — so any library the
review proposes as a replacement must itself clear the 7-axis
Tool-Introduction matrix above. The two sections interlock: the reuse lens
stops reinvention, the tool matrix stops careless adoption, and neither is a
loophole for the other.
Severity guidance
Duplicating an existing in-repo function: MAJOR (two sources of truth
that drift silently).
Reinventing a commodity, well-solved problem as net-new custom logic:
MINOR at L1–L2 blast radius, MAJOR at L3+.
Over-wrapping an adopted library until its guarantees are lost (e.g.
swallowing a library's typed errors behind a bare except): MINOR, or
higher when it defeats a safety property.
A new dependency pulled in to avoid a five-line in-repo helper (dependency
bloat): MINOR, and it still must clear the Tool-Introduction matrix.
Changelog
2026-07-07 (PLAN-153 Wave G, SP-029): added §"Reuse & Prior-Art
Review (reinvented-wheel findings)" — a reviewer lens for flagging
hand-written code that duplicates in-repo utilities, maintained libraries,
MCP capabilities, or sibling skills, interlocked with the existing
Tool-Introduction matrix so neither is a loophole. Clean-room ADAPT merge of
the search-before-build practice; no change to the severity scale, the
blast-radius bands, or the default-BLOCKED verdict discipline.
Skill-Import-Attestation: reviewed-by=AE9B236FDAF0462874060C6BCFCFACF00335DC74; sha256=74ed267ae82ae8fa499b252fd46e9bf3e3292000e85231d4db3b54d458dd8cca
